Zerobox - A Lightweight Cross-Platform Process Sandboxing Tool Based on the OpenAI Codex Runtime

Zerobox is a lightweight process isolation tool that independently extracts the sandbox runtime from OpenAI Codex to prevent risks such as file corruption or data leakage that can occur when executing AI-generated code.

By default, it applies a deny-by-default policy, blocking all writes, network access, and environment variable access that are not explicitly allowed.

Key Features and Characteristics

• Credential Injection: Enables safe external API calls through a network proxy without exposing actual API keys in the code.
• Snapshot and Recovery: Records the state before execution and, after checking changes made during the task, provides the ability to restore immediately.
• Permission Separation: Grants different permissions—such as read-only or specific domain allowlisting—for each LLM tool call, preventing system destruction from prompt injection attacks.
• Supply Chain Attack Prevention: Automatically blocks unintended external network calls during build or test processes to strengthen security.

Technical Specifications

• Platform Support: Supports macOS (Seatbelt) and Linux (Bubblewrap/Seccomp/Namespaces), with Windows support planned.
• Performance and Language: Supports Rust and a TypeScript SDK, and can run with a low overhead of about 10ms without Docker or a VM.